Large-scale Data Analytics for Resilience of Power Grids in Multiple US Service Regions

Transcription

1 Large-scale Data Analytics for Resilience of Power Grids in Multiple US Service Regions Chuanyi Ji (Gatech) Joint work with Yun Wei and Henry Mei (Gatech), Timothy Hays (Central Hudson Electric and Gas), Jorge Calzada and Robert Wilcox (National Grid), Steve Church (New York State Electric and Gas), Brian Nugent and Joe White (Orange & Rockland Utility), Matthew Carey, Gregory Stella and Matthew Wallace (New York State Department of Public Service) July 1, 2016

2 Power Failures in US from Severe Weather > 10 major hurricanes, snow/ice storms in U.S Each caused more than 500,000 customers without power for days Example: Failures during Superstorm Sandy 2012 at the northeast, e.g., New York, Massachusetts,

3 Our Objective Resilience*: Reduce failures, expedite recovery To identify vulnerability through data analytics? Infrastructural vulnerability: Failures Service vulnerability: Recovery *White House report13, Bloomberg report13

4 A bit of background 4

5 Severe Weather Disruptions to Power Grid 90% damages from power distribution [White House 13] Component failures by weather Impact on customers Repairs by crews Distribution Substation Customers Spatiotemporal multi-scale: Minutes, seconds Node, townships, service regions, Distribution Grid 5

6 Challenges Lack of models for spatiotemporally dependent failures, recoveries, and impacts on customers Most prior works are on static models, e.g., Liu et al. 15, Nateghi and Guikema 11 Dynamic models are considered for failures only, e.g., Rudin et al. 12; aggregated recovery by Bertsimas and Mourtzinou 95. Real data need to be detailed and at a large scale A few prior works use real data from one service regions, e.g., Liu et al. 15, Nateghi and Guikema 11, Rudin et al. 12 Aggregated failure data are used at a large scale across US by Larsen et al. 15 6

7 Model guided data analytics 7

8 Non-stationary Spatiotemporal Random Processes Randomness: Failures/recoveries by weather Non-stationarity: Probability distributions vary spatially and temporally Physics: Radial topology and related protection Model: Link a large number of dependent variables on failures, recoveries and impacts

9 Coupled Disruption-Recovery-Cost Processes* Disruption Process: { I[A id (t)] } A id (t): Failed component or activated protective device i Recovery Process { I[D i (v)>t-v] } D i (v): Duration for disruption occurred at v prolonged to t Cost Process { C i (v,t) I[D i (v)>t-v] } C i (v,t): cost from delayed recovery Example cost: Customer interruption time *C. Ji, Y. Wei, R. Wilcox, Large scale data analytics for resilience of power grid across multiple US service regions, Nature Energy, May

10 First-Order Model Parameters Disruption rate: Increment of expected cost/time Conditional probability distribution of delayed recovery given failures Expected cost as the customer outage time

11 Data and Analysis 11

12 Detailed Data from Multiple Service Regions Hurricane Sandy 2012 Upstate NY: ~50,000 square miles, 4 service territories From electric grid ~6600 failures in 2 days Affected ~650,000 customers to 10 days operations in 2012 Details: Failure/min, duration, locations, costs (downtime) on activated protective devices Also data from daily 12

13 80-20 Scaling: 20% failures for 84% affected customers

14 Generalized Scaling Property The scaling property holds for all four DSOs during Sandy DSO: Distribution System Operator

15 Similar Scaling Property for Daily Operations DSO 1

16 Hurricane Exacerbated Vulnerability Probability Daily Operation Hurricane A disruption/minute (~30 times) A top-20% disruption/minute (~170 times)

17 A Cause of Vulnerability? Structure distribution system: Primary, secondary, customer property Locations of top failures: >83% at the primary distribution

18 Infrastructure Vulnerability Local failures can affect tens~hundreds customers Exist in daily operations but exacerbated by Super storm Sandy A cause: How customers are supported by overhead power distribution

19 Customer Downtime: Different from Failures CMI: Customer Interruption Time

20 Small failures matter: Service Vulnerability? A larger number (89%) of small failures (bottom 34% of customers or commonplace devices) amounts to 56% of customer downtime Prioritizing recovery of large failures under available resources does not solve the problem

21 Summary Infrastructural vulnerability A local failure can have non-local impact to customers Exists in daily operations Exacerbated by Super Storm Sandy A cause is the structure of power distribution Service vulnerability: Aggregation of a large number of small failures amounts to major portion of customer downtime Model-guided data analytics shows promise for identifying non-resilience of power distribution and services

22 Data? Collaboration? How to Scale?

23 Reference (see 5 for an extended list) 1 Bertsimas, D. & Mourtzinou, G. Transient laws of non-stationary queueing systems and their applications. Queueing Syst. Theory Appl. 25, , Bloomberg MR. A Stronger, more resilient New York. (City of New York) PlaNYC Reprot, Dobson, I., Carreras, B.A., Lynch, V.E. & Newman, D.E. Complex systems analysis of series of blackouts: Cascading failure, critical points, and self-organization. Chaos: An Interdisciplinary Journal of Nonlinear Science, 17, , Executive Office of the President, Economic benefits of increasing electric grid resilience to weather outages. President's Council of Economic Advisers and the U.S. Department of Energy's Office of Electricity Delivery and Energy Reliability, Washington, DC. Technical Report, C. Ji, Y. Wei, R. Wilcox, Large scale data analytics for resilience of power grid across multiple US service regions, Nature Energy, May Liu, H., Davidson, R.A., David, R.V. & Stedinger, J.R. Negative binomial regression of electric power outages in hurricanes. Journal of Infrastructure Systems, 11, , Larsen, P.H., LaCommare, K.H., Eto, J.H., & Sweeney, J.L. Assessing changes in the reliability of the U.S. electric power system. Report LBNL , R. Nateghi, S. D. Guikema, & S. M. Quiring, Comparison and Validation of Statistical Methods for Predicting Power Outage Durations in the Event of Hurricanes, Risk Analysis, vol. 31, no. 12, pp , December C. Rudin, D. Waltz, R. Anderson, A. Boulanger, A. Salleb-Aouissi, M. Chow, H. Dutta, P. Gross, B. Huang, & S. Ierome, Machine Learning for the New York City Power Grid, IEEE Trans. Pattern Anal. Mach. Intell., vol. 34, no. 2, , Feb Wei, Y, C. Ji, F. Galvan, S. Couvillon, G. Orellana, & J. Momoh. "Learning Geotemporal Nonstationary Failure and Recovery of Power Distribution." Neural Networks and Learning Systems, IEEE Transactions on 25, no. 1, 2014: Executive Office of the President. Economic benefits of increasing electric grid resilience to weather outages. Technical Report, 2013.

24 Incorporating Physical Properties Dynamic Disruption Process Disruption at node i: A id (t), d={f, o}; Damaged/activated component/protective devices Outages induced by failures Failure neighborhood: V (f) i (t) Disruption Process N id (t)=i[a id (t)] Dynamic Recovery Process Recovery event: B (r) i (t), Recovery neighborhood: V (r) i (t) Recovery Process: N ir (t)=i[b ir (t)] Dynamic Neighborhood (Wei et.al.13)

25 Recoveries Expected cost at time t given location z: { } = E S(v) l i f E C(t, z) t ò 0 { ( v S(v) ) E{ G i (v,t) S(v) }}dv ( )dv E{ G i (v,t) S(v) }: Cost by the failures Aggregation over trajectories of failures f l i v S(v) : failures occurred at time v S(v): set of operational nodes at time v { } = EíC i (v,t)p[d i (v) > t - v]+ C j (v,t)p[d j (v) > t - v] E G i (v,t) S(v) ìï îï å jîv i (v) üï ý þï

26 Similar Scaling Property for Daily Operations